A small Paris studio. We design and build production systems on real documents and real tools, backed by an in-house research program on the parts that don't work yet — handwriting that breaks OCR, agents that loop, evals nobody trusts.
Two areas we keep returning to. One because every customer scan looks different; the other because reliability is still the hard part of every agent.
Turning real-world documents into reliable structured data. OCR on handwriting, scans, and low-quality photos. Layout-aware extraction for tables, forms, and non-standard layouts. Long-context understanding grounded in the source with citations.
Tool-using LLM agents built for production. Planners that recover from partial failures, bounded retries that respect token budgets, and replayable evals so reliability is a measured number — not a demo-day vibe.
Three things we hold to across every engagement.
Every change is gated by a written evaluation suite. Numbers move only when we understand why.
Prototypes are wired to real load, real auth, and real latency budgets from day one. No demo-to-prod gap.
We sit inside your team for the duration. Documented handover, paired rollout, then we leave clean.
Have documents or agents worth solving?